Drawing from math I want to say, it depends.
First of all, a text is not generally in that language, but evidence that may allow reconstructing that language. The syllogism calling evidence the fact, the text a language should only be allowable if the language can be reconstructed from the text completely. Of course most texts are too short for that. It may be compared with text that is believed to be in the same language. This biases the comparison; For the worse, I guess, if it can obscure severe mistakes. That is, there is a difference between systematic and spontaneous effects.
Secondly, mutual understanding defines language, so your question seems tautologic in the binary sense. In a gradual sense, your "10%" are arbitrary when not derived from a reasonable, quantifiable measure.
I had just been reading on point-set topology and find that the imagary of a limit of a set and a neighborhood of a limit are a fitting metaphor. One step in the wrong direction, one small mistake too much, can render the message incomprehensible; but there is a known sequence of inferences (the reconstruction) that approach the intended meaning. This is pseudo "abstract non-sense" and my abstract algebra is weak; Started with Closed-form expression and read up-to Open Set. The imagery is Fig 1. Extended with Fig 2.
](https://en.m.wikipedia.org/wiki/File:Neighborhood_illust2.svg)
](https://en.m.wikipedia.org/wiki/File:Runge_theorem.svg)
Another problem is that your 99% are supposed to be competent speakers, but this property cannot be established without the sought metric. It leads to a slippery slope well known in the categorization of e.g. the development from Latin to Italian, if Dante is the proto-type for not-Latin. And it leads to arguments about how you can be sure that you understand the devious text correctly (which is a general problem).
For a German English speaker, it sounds as if you described Dutch, though the error rate is closer to 100% percent and my understanding thus rather limited unless I know what's being said. You might need to weigh the defective features for an accurate percentage, in which case the idioms, grammar morphology and verbal base should appear much closer to 10% defective (the numbers arw figurative, orders of magnitude). I know you said no varieties but I like to think of Dutch as Rudi Karell's accent when speaking German, and any real written Dutch, read with my own intonation just appears as an immense speech impediment. (No offence intended, I love it). The whole "dialect with an army" idea applies. But note that boundaries by definition can be approached from both sides.